Scholarlink

<p>Visible–infrared person re-identification (VI-ReID) faces the primary challenge of extracting discriminative features from different modalities to achieve effective matching. Existing methods primarily focus on minimizing modality differences but often fail to fully exploit the potential of modal information due to limitations such as modality diversity and limited sample quantities. To address this issue, we propose the original–channel contrastive learning network (OCCLNet), which aims to expand dataset information to contrast positive and negative samples, thereby reducing the modality gap. OCCLNet introduces channel augmentation techniques to balance the information disparity between RGB and IR images and proposes the channel dimension enhanced (CDE) module to strengthen the discriminative information of person images. Furthermore, by constructing novel channel–original-level positive–negative sample contrastive strategies (COCS), the model learns common features among positive samples and distinguishes differences from negative samples in the high-level feature space. We also design a supervised center contrast aggregation loss (CCL) that incorporates the concept of feature centers into positive and negative sample contrasts, reducing intra-class variance and increasing inter-class differences. Extensive experiments on datasets such as SYSU-MM01 and RegDB demonstrate the outstanding performance of OCCLNet, achieving state-of-the-art recognition accuracy. We have distributed the code on <a href="https://github.com/Hua-XC/OCCLNet">https://github.com/Hua-XC/OCCLNet</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing visible–infrared person re-identification via original–channel contrastive learning

  • 期刊论文

  • 发表日期：

    2025年9月12日

查看全文
  • Xuecheng Hua,
  • Gege Zhu,
  • Ke Cheng,
  • Shitong Wang

摘要

Visible–infrared person re-identification (VI-ReID) faces the primary challenge of extracting discriminative features from different modalities to achieve effective matching. Existing methods primarily focus on minimizing modality differences but often fail to fully exploit the potential of modal information due to limitations such as modality diversity and limited sample quantities. To address this issue, we propose the original–channel contrastive learning network (OCCLNet), which aims to expand dataset information to contrast positive and negative samples, thereby reducing the modality gap. OCCLNet introduces channel augmentation techniques to balance the information disparity between RGB and IR images and proposes the channel dimension enhanced (CDE) module to strengthen the discriminative information of person images. Furthermore, by constructing novel channel–original-level positive–negative sample contrastive strategies (COCS), the model learns common features among positive samples and distinguishes differences from negative samples in the high-level feature space. We also design a supervised center contrast aggregation loss (CCL) that incorporates the concept of feature centers into positive and negative sample contrasts, reducing intra-class variance and increasing inter-class differences. Extensive experiments on datasets such as SYSU-MM01 and RegDB demonstrate the outstanding performance of OCCLNet, achieving state-of-the-art recognition accuracy. We have distributed the code on https://github.com/Hua-XC/OCCLNet.